LLMs / financial data
alphasig
I work on alphasig, a Python tool that extracts structured information from SEC filings. My contributions include multi-hop supply-chain queries, incremental extraction, webhook delivery, and improvements to signal scoring and reliability.
Handling a first filing
The tone engine discarded a company’s first filing because it had no earlier filing to compare against. I added a separate baseline result, explicitly marked so it could be distinguished from a change in tone. I also added regression tests for informative and neutral baselines.
Accounting for uncertainty
Risk changes were scored by severity alone, with a fixed confidence value. I added model-reported confidence to the extraction schema and used it to weight signal strength. An uncertain severe change no longer receives the same score as one the model reports with higher confidence.
Pipeline and graph features
I added queries for multi-hop supplier and customer dependencies, date-filtered incremental extraction, signal deduplication, and webhook delivery. I also expanded retries to cover connection failures, timeouts, and server errors, and added validation-error logging.